- Modeling financial decisions over space and time
- Local measurement of macroeconomic variables
- House price dynamics
- The value of innovation to companies and the associated risks for workers
- (Spatial) labor market risk and implications for asset prices
- Understanding the distributional effects of policies, modeling heterogeneity
- Modeling missing data & representing uncertainty
- Bayesian empirical asset pricing
CVaR Modeling with Mixed Frequency Data
Joint with Kevin Duncan
Conditional value at risk modeling is extremely important for modeling downside events, like negative shocks to portfolios of financial assets or loan default risks. We propose a method for incorporating infrequently measured variables which may be important in predicting downside risk in the same model as high-frequency observations.
Credit Constraints and House Price Dynamics
Status: Early stage
I’m exploring how and whether credit constraints drive house price dynamics. Very early stage, hopefully posts to follow!
LATE to Great (Link)
This was my writing sample for grad school applications. The basic idea is to estimate treatment effect heterogeneity with very small subgroups by putting weak structure on the relationships across cells (in this case based on neighbor relationships on a lattice constructed over the covariate space). Relates to much better work on GMRF priors by Gao, Kennedy,Simpson, and Gelman and Gao, Kennedy, and Simpson. I discuss the structure of the priors and show in simulations that it can recover the shape of treatment effects over a covariate space very accurately even when subgroups are quite small. Further, I show an application to 2SLS; without structural priors 2SLS approaches infinite variance as the number of cells groups, but with structural GMRF priors this is no issue at all, even when the number of cells exceeds the number of data points.
Modeling Missing Data with Nonlinear Quantile Regression
Joint with Jon Rothbaum, Larry Schmidt
Missing data is a large problem in modern surveys, and a substantial subset of important variables lack any well-known parametric form. We relax parametric assumptions by using the quantile spacings approach of Schmidt and Zhu, interpolating the fitted quantiles with splines to generate a predictive distribution.
Inference and Prediction of Stock Returns using Multilevel Models (Link)
Joint with Sam Thomas
This is a paper from when I was first learning about Bayesian hierarchical models that I wrote during my time at Capital Group, a large active investment firm. The gist is that if you embed Fama-French style regressions inside a hierarchical model with stock-level interactions you get far better risk forecasts out of sample (sort of a full-Bayes version of the insights in Vasicek 1979).
R package making it easy to run Bayesian meta-analysis models with Stan on the back-end.
Implements the quantile spacings approach of Schmidt and Zhu, modeling the log distance between quantiles via quantile regression rather than a linear quantile model as a simple way to prevent quantile crossing.
Massachussetts Institute of Technology
University of Texas at Austin
Subject: Music, Economics
Software: R, Python, C++, Stan, Stata
Music: piano, saxophone, guitar (kind of), arranging & composition